BRIDGING DATA AND ELECTRONIC DANCE MUSIC

by

Jasdev Singh

Submitted in partial fulfillment of therequirements for the degree of

Bachelor of Science Computer Science

at

University of VirginiaCharlottesville, VA

March 2014

c© Copyright by Jasdev Singh, 2014

UNIVERSITY OF VIRGINIA

DEPARTMENT OF COMPUTER SCIENCE

The undersigned hereby certify that they have read and recommend to the Faculty

of Undergraduate Studies for acceptance a thesis entitled “BRIDGING DATA AND

ELECTRONIC DANCE MUSIC” by Jasdev Singh in partial fulfillment of the

requirements for the degree of Bachelor of Science Computer Science.

Dated: March 9, 2014

Supervisor:Dr. Yanjun Qi

Reader:Matthew N. Eisler, PhD

ii

UNIVERSITY OF VIRGINIA

DATE: March 9, 2014

AUTHOR: Jasdev Singh

TITLE: BRIDGING DATA AND ELECTRONIC DANCE MUSIC

DEPARTMENT OR SCHOOL: Department of Computer Science

DEGREE: B.S. CONVOCATION: May YEAR: 2014

Permission is herewith granted to University of Virginia to circulate and to havecopied for non-commercial purposes, at its discretion, the above title upon the requestof individuals or institutions. I understand that my thesis will be electronicallyavailable to the public.

The author reserves other publication rights, and neither the thesis nor extensiveextracts from it may be printed or otherwise reproduced without the author’s writtenpermission.

The author attests that permission has been obtained for the use of any copyrightedmaterial appearing in the thesis (other than brief excerpts requiring only properacknowledgement in scholarly writing), and that all such use is clearly acknowledged.

Signature of Author

iii

Table of Contents

Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v

Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi

Chapter 1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.1 Background on dance music . . . . . . . . . . . . . . . . . . . . . . . 1

1.2 Motivation for study . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

Chapter 2 Previous work . . . . . . . . . . . . . . . . . . . . . . . . . 4

2.1 Research in EDM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2.2 Research in sentiment analysis . . . . . . . . . . . . . . . . . . . . . . 4

Chapter 3 Overview of data sources . . . . . . . . . . . . . . . . . . . 6

3.1 Dataset 1 - DJ Magazine Top 100 . . . . . . . . . . . . . . . . . . . . 6

3.2 Dataset 2 - 1001Tracklists . . . . . . . . . . . . . . . . . . . . . . . . 7

3.3 Dataset 3 - SoundCloud comments . . . . . . . . . . . . . . . . . . . 8

Chapter 4 Social Dynamic Analysis of DJ Magazine Top 100 . . . 9

4.1 Top 100 Case Study #1, Swedish House Mafia . . . . . . . . . . . . . 9

4.2 Top 100 Case Study #2, Trance rankings . . . . . . . . . . . . . . . . 10

4.3 Top 100 Case Study #3, Mentor rankings . . . . . . . . . . . . . . . 10

4.4 Top 100 Case Study #4, Record labels . . . . . . . . . . . . . . . . . 11

Chapter 5 Social network analysis of festivals and live events . . . 12

5.1 Electric Zoo Highlights, opportunities for breakout DJs . . . . . . . . 13

5.2 Electric Zoo Highlights, Sets with Isolated Tracks . . . . . . . . . . . 14

5.3 Electric Zoo Highlights, most played artists . . . . . . . . . . . . . . . 15

Chapter 6 SoundCloud Sentiment Analysis . . . . . . . . . . . . . . 17

6.1 Data mining overview . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

iv

6.2 Training process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

6.3 Tools used . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

6.4 Best window algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . 19

6.5 Visualizations and extensions . . . . . . . . . . . . . . . . . . . . . . 19

6.6 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

Chapter 7 Conclusion and Future Work . . . . . . . . . . . . . . . . 22

Chapter 8 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

v

Abstract

Electronic dance music (EDM) is a relatively new genre with unique features such as

a large number of subgenres and the notion of playing live events in a festival format.

With its rise to mainstream popularity, there are opportunities to explore the social

dynamics of EDM by analyzing datasets associated with the genre. In our research,

we seek to retrieve information from three such datasets, the DJ Magazine Top 100,

live festival set lists, and SoundCloud comment data. The data mining techniques

presented in this thesis provide great potential for helping listeners, musicians, and

online storefronts, e.g. automatic recommendation and music summarization.

vi

Acknowledgements

• Dr. Yanjun Qi, PhD, Department of Computer Science

• University of Virginia Association of Computing Machinery

vii

Chapter 1

Introduction

1.1 Background on dance music

Electronic dance music (often called EDM) includes multiple electronic music genres,

usually for dance-based events (festivals and clubs). At these live events, it is very

common for disk jockeys (DJs) to play continuous sets, ranging from 30 minute to

multiple hour mixes. Moreover, producers and DJs will often collaborate on and play

each other’s music. Below is a graph from a recent report by Google Music’s Research

Group1 showing the popularity of various genres uploaded to the Google Play store2,

since the service’s inception. Highlighted by the red box is Dance and Electronic

music. It can be seen that EDM is roughly the 6th largest genre, as of the most

recent year in the survey.

Figure 1.1: Google Music genre timeline - rise of dance music

A unique feature of dance music is the wide array of subgenres present. This allows

for the rise of communities of producers and opportunities for remixing originals into

other subgenres. Zooming in on the graph presented previously, we can see the

diversity of the overall genre, with notable examples being trance, techno, house,

ambient, and drum & bass.

1

2

Figure 1.2: Subgenres present within Dance music

Next up, we will highlight some of the most popular genres within EDM. To be-

gin, House is a genre typically identified by a 4 by 4 kick drum and a 120 - 130 beats

per minute (BPM) range. Some common names within House are Avicii, Swedish

House Mafia, and David Guetta. While bucketing artists into genres is not an exact

science, one can generally associate artists with genres based on the majority of their

productions. Secondly, Electro is a style with identifiable synths and a tempo range

of 125 - 135 BPM. Some notable artists in the space are Dada Life, Justice, and Boy’s

Noize. Third is Dubstep, defined by an oscillating wobble bass. The BPM range hov-

ers around 140 and notable producers in the genre are Skrillex, Bassnectar, Skream,

and Benga. Lastly, Trance is another large subgenre popular around the globe. Its

tempo range is wide, 125 - 150 bpm, and is noted by anthems and exaggerated builds.

Trance has been doing extremely well in the DJ Magazine Top 100, which we will

introduce soon, and is lead by artists such as Armin van Buuren, Above & Beyond,

and Ferry Corsten.

1.2 Motivation for study

Being a relatively young genre, there has not been a lot of research done on the EDM

social demographic and artists. In this paper, we will take a look at three datasets

to gain information about this style of music. To start, we will note trends in the

DJ Magazine Top 100 poll, which is an attempt to put a stack rank on DJs and

producers from year to year. Insights on this chart can help us to better understand

3

dynamics such as the rise and fall of genres and record labels. Secondly, we will look

at set lists of what DJs played at live events, specifically one festival in New York

called Electric Zoo. From these set lists, we can make connections from specific DJs

to others by how they played and/or remixed each other’s work. Such connections

naturally lead to the formation of a network diagram. From such diagrams, we can

perform an analysis on communities present to gain insight on the underlying social

constructs. Lastly, we will perform a sentiment analysis on public user comments on

music hosted on the popular online streaming service, SoundCloud. From the text

data housed in these timestamped comments, we will attempt to form a heuristic

for the overall positivity towards the underlying audio at that portion of the music.

This is hugely beneficial for two major reasons. First, given any song, we can now

attempt to answer the question of ’what is the best portion of this song?’ via the text

comments as a proxy for the audio. Extending this, an answer to such a question can

help in summarizing larger pieces of audio or even serve as the recommended preview,

prior to buying music on online storefronts. To recap, our motivating factors are as

follows:

• Better understand the social dynamics between entities in the genre

• Retrieve information about musical quality via sentiment analysis

• Construct a baseline ’best window’ algorithm to be used by online storefronts

This research will help better inform artists and musicians about how their work

fits in the larger network of Electronic Dance Music as a whole. Additionally, listeners

and online music storefronts can benefit from our ’best window’ algorithm developed

in this paper to preview tracks and better present content to users, respectively.

Chapter 2

Previous work

2.1 Research in EDM

Previous work in the realm of analyzing data on EDM is far and in between. However,

Eventbrite, an online ticketing platform for live events, put out a study in June 2013

showcasing behavioral trends on how EDM listeners shared content on social media

before and after events3. This serves as a perfect contrast to our study of how DJs

play music at live events, giving us a holistic view. Some of the key takeaways from

the study were:

• EDM fans are twice as likely as other music fans to want to attend an event

when their friends post on social media

• Over half of EDM fans revealed that they would pay for live streaming of an

EDM concert

• People who attend EDM events are significantly more likely to share on social

media before (67%), during (41%) and after (63%) an EDM event than people

who attend music events but not EDM events (with 37%, 21%, 51% respectively)

2.2 Research in sentiment analysis

Sentiment analysis is a branch of Natural Language Processing that we will rely

heavily upon in our attempt to gain insights from text data present on the music

streaming service, SoundCloud4. More specifically, sentiment analysis attempts to

determine the attitude and polarity of a document in both supervised and unsuper-

vised contexts. In this paper, we will employ a supervised machine learning technique

with a training corpus of thousands of SoundCloud comments to detect positive and

negative emotions. As a field of study, sentiment analysis has gained popularity with

4

5

the rise of social media, as the amount of reviews, ratings, and recommendations have

proliferated.

Chapter 3

Overview of data sources

Figure 3.1: Three datasets in which we will perform data mining - DJ Magazine Top100 (left), 1001Tracklists (middle), SoundCloud (right)

3.1 Dataset 1 - DJ Magazine Top 100

Using data available within the community, we seek to convert it into knowledge using

a variety of analysis techniques. To begin, we are going to use poll data from the DJ

Magazine Top 100. Starting in 1997, DJ Magazine has put together a popular vote

poll to determine the top 100 DJs each year. However, when it comes to subjective

topics, like music, this type of poll has received criticism since popularity does not

always coincide with musical talent. But, on a higher level, we can use this time series

data to draw larger and more general conclusions about the genre and its constituents.

In our study, we scraped the poll data and made it available in both CSV5 and JSON6

format. Some of the challenges faced when scraping this data include:

• Resolving accented characters in DJ names

• Naming consistency resolution (e.g. ’DJ Snake’ versus ’Snake’)

6

7

• Artist stage name changes (defaulted to latest known stage name)

3.2 Dataset 2 - 1001Tracklists

Second, we will be looking for connections between DJ set lists that are available

on the popular site, 1001 Tracklists. As mentioned before, it is common for DJs to

play longer sets during live events. This site serves as an aggregator for the tracks

played in such sets to better allow for song discovery. Utilizing this, we will scrape the

tracklist data for a specific event, Electric Zoo NYC7, and try to establish connections

between artists by the music they play in each others’ sets. This can allow for the

creation of a community diagram, where we can logically see higher level connections

between groups of DJs. In parsing this data, JSON was used in a manner reminiscent

of an adjacency list, typically used in graphical settings. See below for a sample entry:

Figure 3.2: Sample schema for describing social networks among DJs

In the figure above, the artists in question is Mat Zo, as indicated by the name

key. Ranking information such as the average rank and size are used to scale the

nodes in the diagram. Lastly, the played key maps to an array of other unique artist

names, whose work appeared in Mat Zo’s set at the festival. Using this schema, we

can generate a network diagram of artists at an event, to attempt to gain information

about underlying group dynamics.

8

3.3 Dataset 3 - SoundCloud comments

Lastly, we will mine text data from the music streaming service SoundCloud. The

motivation for this inquiry is the fact that comments on the service are timestamped

at specific points during songs. So, using sentiment analysis, we can seek to get a

better understanding about the underlying audio of a track, given what people are

saying in text. Our initial data mining fetched over 200,000 of these comments and

thousands of them were labeled manually as being either negative, neutral, semi-

positive, or really positive in emotion. These labeled points were then used to train

a classifier to distinguish sentiment on new unlabeled comments.

Chapter 4

Social Dynamic Analysis of DJ Magazine Top 100

4.1 Top 100 Case Study #1, Swedish House Mafia

To exemplify some of the trends in these charts, we begin with the DJ trio that

is Swedish House Mafia. Starting in 2008, Steve Angello, Sebastian Ingrosso, and

Axwell began touring around the world together and recently stopped in 2013. Below

is a plot of their respective rankings in the chart, including the ranking of the group

overall:

Figure 4.1: Temporal rankings of Swedish House Mafia Members

From this figure, we can see that upon the group’ splitting in 2013, both Sebastian

Ingrosso and Axwell had higher rankings in the poll, even above the group, whereas

Steve Angello’s regard in the public eye was ranked lower. The reasoning for the

group’s spit is unknown (although a documentary http://leavetheworldbehind.

com/ is set to be released soon), but this may help us gain some insight on the

situation.

9

10

4.2 Top 100 Case Study #2, Trance rankings

Trance is an extremely popular electronic genre that has typically held a strong

showing in the polls. However, in 2013, there has been some pushback from fans

against the genre, as DJs and producers have been experimenting with more ”radio”

style tracks. Below we have a plot of all the trance DJs appearing in the Top 100:

Figure 4.2: Temporal rankings of all DJs related to the Trance genre

In 2013, every Trance DJ in the poll, with the exception of Bobina and Andrew

Rayel, actually declined in ranking, when compared to the previous years.

4.3 Top 100 Case Study #3, Mentor rankings

Figure 4.3: Temporal rankings dynamics of a pair of mentor/mentee DJs

11

Alesso, a 22 year-old DJ from Sweden, got an early start in 2010. Later on in

his career, he received mentoring from the aforementioned Sebastian Ingrosso. This

mentor / mentee relationship has been interestingly reflected in the Top 100 charts.

From the graph below, you can see that Alesso’s rank actually surpassed Ingrosso’s,

despite him being the mentee.

4.4 Top 100 Case Study #4, Record labels

Like other genres, it is common in EDM for artists to come together in groups to

form record labels. Applying this grouping to our Top 100 data, we can see the

performance of various labels in the poll.

Figure 4.4: Temporal rankings dynamics of record labels

These graphs can help us see how the label is performing overall in the charts and

identify integral artists.

Chapter 5

Social network analysis of festivals and live events

Large scale EDM events have become increasingly common, with upwards of 300 thou-

sand attendees in some cases. During such events, DJs will often play sets composed

of not only their work but work from other artists as well. This scenario lends itself

well to a sort of network analysis. To get tracklist data, there is a crowd-sourced site,

1001Tracklists that hosts set lists from a wide array of events and podcasts. More-

over, users can go in and attach audio samples to tracks, link unknown songs to other

appearances in different contexts, and even resolve disputes regarding an unknown

song via voting. To use this data, we scraped the site into JSON and crossed it with

Top 100 rankings to get a bigger picture of community dynamics within festivals. At-

tempting to showcase this, we decided to use data from an annual festival run in New

York, Electric Zoo. Below is a graph of artists present in the festival, with the font

of their names scaled by their respective Top 100 ranking. Moreover, a black colored

label is the current source DJ, with green edges representing the origin DJ playing a

song by the source DJ and the red edges represent that the source DJ played a song

by the destination DJ. An interactive version of this diagram is available online8.

12

13

Figure 5.1: Social network diagram of DJs at a festival event called Electric Zoo

5.1 Electric Zoo Highlights, opportunities for breakout DJs

A great opportunity for breakout artists at these events is the scenario where a highly

ranked DJ plays one to their songs. This is desired because headliners usually have

slots on bigger stages, when compared to younger artists, who typically perform on

side stages with a smaller audience. To highlight such a scenario during Electric Zoo,

Avicii played a song by the young duo Dog Blood.

14

Figure 5.2: Example of a breakout artist (Dog Blood) being played by an establishedDJ (Avicii)

As noted in the diagram, Avicii’s average rank in the charts is 12.75, whereas Dog

Blood has never actually appeared on the poll. This edge marks a huge opportunity

for the duo to gain an audience and following.

5.2 Electric Zoo Highlights, Sets with Isolated Tracks

Moreover, with this diagram, we can notice specific DJs who are disjoint in con-

nections to any other DJs present in the festival. Below, John O’Callaghan, Danny

Tenagila, and Fehrplay all had set lists that were disjoint from others in the festival.

This is likely attributed to their genres being relatively niche.

15

Figure 5.3: Social Network Diagram showing artists with disjoint sets in a live event

5.3 Electric Zoo Highlights, most played artists

Lastly, we can seek to answer the question of which artist was most played at the

festival. This is calculated by iterating over all artists and keeping track of the one

with the largest number of incoming edges. From Electric Zoo, Knife Party had the

most appearances in other sets with 12 plays.

16

Figure 5.4: Social network diagram showing the most played DJ (Knife Party) at alive event

Chapter 6

SoundCloud Sentiment Analysis

6.1 Data mining overview

To begin training an accurate classifier in determining sentiment, we need a lot of

training data. Using SoundCloud’s public REST API9 we wrote a script to fetch

200,00 comments on songs in the EDM genre. These were stored locally in flat text

files for labeling.

6.2 Training process

Prior to manually labeling the comments collected, a few steps were taken to clean

the data. First, an unfortunate part of the SoundCloud service is the presence of

spammers who tend to comment with links to his/her music, to gain views. Such

data would interfere with our core sentiment analysis, so all comments with URLs

were stripped from parsing. Additionally, non-English comments were removed to

specifically focus on deriving sentiment in the English context. Lastly, all text was

converted to lowercase and stop words (words that do not add sentiment value such

as ’the’, ’is’, etc.) were removed. A special note about the stop words is that both

a standard list of English stop words and a custom stop word list were used. The

English stop word list was taken from the default module provided in the Python

Natural Language Toolkit (NLTK). The need for an additional custom stop word

list is because in EDM there are common words, used in context, that do not add

sentimental value. Examples of these words are ’remix’, ’listening’, and ’tune.’

After this cleansing process, the comments were then manually labeled by splitting

them across four text files, representing negative, neutral, semi-positive, and really-

positive comments. The logic behind this range scale as opposed to a binary scale

is that sentiment often falls along a gradient as opposed to being clearly positive or

17

18

negative. Splitting the comments up in separate text files allowed us to load them

into memory later, labeled by their respective filename.

Figure 6.1: System diagram of how SoundCloud comment data was fetched andclassified. Step 1: Given a list of hundreds of SoundCloud URLs, fetch their respectivecomments and save locally in a flat text file. Step 2: Manually label each commentalong the scale of negative, neutral, semi-positive, or really-positive according tosentiment

6.3 Tools used

To perform our sentiment analysis, we used a Naive Bayes Classifier from Python’s

Natural Language Toolkit10. This classifier takes in our labeled comments from above

and creates a frequency table of common tokens and their appearance in specific

labels. For example, the token ’great’, appeared in many really-positive comments,

so when a new comment is introduced to the classifier, the presence of the word ’great’

will likely bump the comments label towards being really-positive. Below is a sample

of some of the informative features of the classifier:

Token Label Ratio Value

bad negative : semi positive 15 : 1

holy really positive : semi positive 15.1 : 1

beautiful really positive : semi positive 11.2 : 1

best really positive : semi positive 7.9 : 1

nice semi positive : really positive 6.8 : 1

stop neutral : semi positive 4.3 : 1

Table 6.1: Top 6 informative features of the Bayes Classifier and the associated label

ratio values

19

6.4 Best window algorithm

With our trained classifier at hand, we can begin our analysis of SoundCloud songs.

Below is a diagram of how comments typically show up during playback:

Figure 6.2: How text comments are rendered on SoundCloud tracks

Using these comments, we seek to answer the question of what window of this

song is the ’best’ in terms of user sentiment, as derived from the text comments. We

can accomplish this in the following steps:

1. Given a SoundCloud URL, use the REST API to fetch all timestmped comments

associated with the track

2. If we seek to find the best 10 seconds, we create a sliding 10 second window

across the duration of the track

(a) For each 10 second window, we score each comment in the window accord-

ing to its classified label outputted from the Bayes Classifier. Attributing

a score to each label (e.g. negative is -1, neutral is 0, semi-positive is +1,

and really positive is +2). Summing these weights we get a score for the

entire window.

3. Keep track of the window with the highest score from above loop and output

as ’best’ portion of track.

6.5 Visualizations and extensions

Running the algorithm, we can create a heat map of how comment sentiment is

distributed along a given track.

20

Figure 6.3: Heat map of sentiment of SoundCloud comments. Red region indicatedthe portion of the track with the highest overall sentiment score

Above, the lines represent user comments and the closer the comment is to the

color green, the more positive it is in sentiment. Additionally, the window in red

represents the best 8 second window of the song via the algorithm above. To extend

this notion of extracting the best parts of songs, we took 10 of the top 100 songs on

Beatport’s Top 10011 and ran this algorithm on them. From these 10 best portions,

we created a single track composed of those best parts12. Lastly, from the text data,

we can also visualize the comments in for form of a word cloud, where the font of a

token is scaled by frequency.

Figure 6.4: Word cloud of tokens present in comment text data. Font is scaled byfrequency of token in comment corpus.

6.6 Results

We report the Precision/Recall score to evaluate our experiments.

Precision: Among the negative (neutral or semi-positive or positive) comments de-

tected by the system, how many are true negative comments ?

21

Recall: For all possible negatives (neutral or semi-positive or really-positive), how

many were detected by system?

A good detector should have both high precision and high recall. If we let d be the

number of negative (neutral or semi-positive or positive) detected, z be the number

of negatives manually judged, and dz be the number of negative manually judged as

negative among the detected ones, the precision and recall for the system are:

precision =dz

d(6.1)

recall =dz

z(6.2)

Comment corpus breakdown

Negative Neutral Semi-positive Positive

Total Count: 133 114 493 425

Percentage of Corpus: 11.4 9.8 42.3 36.5

Precision Rate: 1.0 1.0 0.601 0.932

Recall Rate: 0.083 0.088 0.982 0.744

Error Rates - Cross validation

2-Fold 5-Fold 10-Fold

Error Rate: 0.146 0.167 0.423

Chapter 7

Conclusion and Future Work

From the three datasets analyzed above, we can better explore the trends and con-

nections between the main actors within the EDM industry. From the DJ Magazine

Top 100, the temporal nature of the popularity of genres is revealed as the rise and

fall in the charts. Additionally, there are dynamics between artists, as set lists from

Electric Zoo demonstrate that there exist communities of DJs who play each others

music. Lastly, from analyzing text data from SoundCloud, we can reach a heuristic

for the ’best’ portion of a song via sentiment analysis on timestamped data points.

The last section, where we developed the ’best window’ algorithm, as a heuristic

to determine the portion of a track with the best overall sentiment score, is extremely

valuable for both listeners and online music storefronts. Listeners can use this to bet-

ter browse musical libraries by listening to trimmed previews of longer tracks via this

algorithm. This has implications in content discovery, as users can now find music

in a faster manner, by listening to key parts of music. On the online storefront side,

companies can use sentiment data to effectively make better previews for content that

are presented to customers before purchase.

Being a relatively young genre, there are many applications and extensions to the

research presented. For instance, the best window algorithm can be used in contexts

such as summarization of audio. Moreover, when online storefronts present previews

of tracks, they can use the best windows as a default choice for clips to show users

prior to purchase.

On a final note, all of the code and work from this project is open source13.

22

Chapter 8

References

1 http://research.google.com/bigpicture/music/

2 https://play.google.com/store/music

3 http://www.eventbrite.com/pressreleases/edm-fan-behavior-differs-greatly-

from-that-of-other-music-fans/

4 https://soundcloud.com/

5 https://github.com/Jasdev/soundcloud-sentiment/blob/master/heatmap-gen/

static/datasets/dj-mag-top-100.csv

6 https://github.com/Jasdev/soundcloud-sentiment/blob/master/heatmap-gen/

static/datasets/dj-mag-top-100.json

7 http://www.madeevent.com/ElectricZoo/

8 http://www.jasdev.me/thesis/ezoo

9 http://developers.soundcloud.com/

10 http://www.nltk.org/

11 http://www.beatport.com/top-100

12 https://soundcloud.com/jasdev-singh/beatport-top-10-drops

13 https://github.com/Jasdev/soundcloud-sentiment

23